There’s some cool things happening with AI bots. But what about when it comes to making video games? Turns out, AI can do that too.
Nvidia has unveiled GameGAN, a generative adversarial neural network (GAN) that’s capable of generating a fully-functional version of PAC-MAN without the help of an existing video game engine. The network was trained with 50,000 episodes of Pac-Man provided by Bandai Namco, and is the first attempt to emulate a game engine using a GAN.
GANs are basically two neural networks in one, with the networks competing against each other. As the networks continually clash over thousands and thousands of episodes of gameplay, they continually learn about how the video game is supposed to function. The constant repetition trains the networks to visually imitate the target game, without having direct access or data from the game itself. Over time, the model learns the difference between static and dynamic elements, where Pac-Man can and can’t go, what happens when he runs into a ghost, and so on.
“We wanted to see whether the AI could learn the rules of an environment just by looking at the screenplay of an agent moving through the game. And it did,” Seung-Wook Kim, Nvidia researcher and lead on the project said.
GameGAN was trained on a modified version of Pac-Man, but also VizDoom. VizDoom is the fork of id’s classic shooter that’s used for deathmatch tournaments between AI agents.
[referenced url=”https://www.kotaku.com.au/2016/04/theres-a-new-doom-competition-but-its-for-ai-only/” thumb=”https://www.kotaku.com.au/wp-content/uploads/sites/3/2016/04/doom-original-3-410×231.png” title=”There’s A New DOOM Competition, But It’s For AI Only” excerpt=”The internet has been privy to the results of StarCraft bots battling each other for years. Google’s AlphaGo AI recently conquered a game many thought once impossible. Bots playing games are cool, and programmers like making them. But there hasn’t been a competition for bots playing DOOM — the original DOOM, that is. Fortunately, Visual Doom AI plans to fix that.”]
So what this means in practice is that developers – or anyone who wants to mess around with it, really – can use the GameGAN AI to create new levels or layouts, since the networks can be trained on any game as long as they’re given footage. The principle has a lot of use outside of games too, if you think about the possibilities of an AI that can learn actions or the rules of an environment just through observation.
Nvidia’s AI-made version of Pac-Man will be released later this year through their AI Playground portal. Nvidia’s research will also be presented to the Conference on Computer Vision and Pattern Recognition conference later this year, although you can read the paper for yourself here.
Leave a Reply